Generalized Predictive Control Method Based On Latent Variable Space

Along with increasing of global industrialization and expanding of manufacturing scale,the technological flow have become more and more complicated in many large-scale industrial processes which usually exist the characteristics of multivariable,strong coupling and nonlinear,these make the data-driven modeling methods become a hot research area gradually.Partial least squares(PLS)which as a multivariate statistical method,has been widely used in fault diagnose,soft sensor and other fields,however,in control field,the research of PLS is still in early stage.This paper which integrates PLS with the MPC,researches several kinds of generalized predictive control method based on latent variable space,the main contents include:Aiming at the multivariable and collinearity existed in practical industrial process,a generalized predictive control method based on ARX-PLS dynamic model is introduced.This method makes use of the PLS to build the latent space,where the coupling and collinearity of process can be eliminated,so the original multivariable control problem can be decomposed into several univariate ones in parallel.The predictive controller is simplified and the computing time is reduced by the proposed method.Simulation studies with quadruple tank process are presented to demonstrate the control effect.Next a new dynamic modeling method is proposed which makes an improvement on the basis of the ARX-PLS dynamic modeling,namely ARX-Dynamic inner PLS model.This method not only describes an explicit dynamic inner model in latent space,but also keeps consistent between the inner with the outer model at the same time.Then the generalized predictive controller is designed.Aiming at the model mismatch caused by environmental factors,we combine the recursive least squares(RLS)and recursive partial least squares(RPLS)to realize the adaptive update of inner and outer model parameters,which guarantees the real-time matching between model and object.Simulating studies verify the effectiveness.A generalized predictive control method based on local LSSVM model in latent space is proposed to deal with the problem associating with time-variation and strong nonlinearity in industrial process systems.To begin with,the latent space is constructed by PLS,subsequently,the relevant data samples at present are selected out by Just-in-time learning(JITL),which are utilized to build the local LSSVM model of each SISO subsystems online in latent space,finally,the generalized predictive control is implemented to these subsystems separately.Simulation studies with quadruple tank and pH neutralization process are presented to verify the effectiveness.An experiment study for process control device has been presented.First we use the OPC and the configuration software to realize the communication between MATLAB and the device,and then the predictive control method based on latent space can be applied into the practical device,which shows the actual control effect.